Research of the VQ-f16 VAE latent space compression methods for FPV video stream

The efficiency of video stream transmission links between an unmanned aerial vehicle and its operator in mobile and hybrid orbital-terrestrial communication networks directly depends on solving the problem of compressing video stream frames, while ensuring the quality of the restored image. One of the methods of frame compression is the use of variational autoencoders to transfer the latent space obtained during the processing of individual frames. The present paper is devoted to the research of how effectively different algorithms can compress quantized latent space of variational autoencoder VQ-f16 from Stable Diffusion repository. The system of quantized latent space compression algorithms efficiency indicators and their description are presented. A comparative analysis of the efficiency of quantized latent space compression algorithms is conducted. The results of analyzing the efficiency of quantized latent space compression algorithms are presented and recommendations for improving their efficiency are given.